Decision theory and the "almost implies near" phenomenon

We examine behavioral axioms in decision theory that are satisfied approximately rather than exactly. We demonstrate that in key domains – decisions under risk, uncertainty, and intertemporal choice – behavior that \emph{almost} satisfies an axiom implies the existence of a utility function that is \emph{near} one that adheres to the standard theoretical representation (e.g., expected utility, or exponentially discounted utility). We explicitly quantify the distance between the utility that captures actual behavior and the ideal theoretical utility as a function of the measured deviation from the axiom. This result formally connects two distinct quantitative exercises: measuring empirical deviations from theory and utilizing approximate optimization. Effectively, we show that small deviations from behavioral axioms rationalize the use of standard models as valid approximations.

💡 Research Summary

The paper “Decision theory and the ‘almost implies near’ phenomenon” investigates what can be said when the fundamental axioms that underlie the standard representations of decision theory are only approximately satisfied. Rather than assuming that an axiom A holds exactly, the authors introduce an ε‑relaxed version A(ε) in which the axiom may be violated by a small, observable amount ε. Their central claim is that if A(ε) holds for a preference relation, then there exist two suitably normalized utility functions u and v such that: (i) u exactly represents the observed choices (the “behavioral” utility); (ii) v satisfies the exact axiom A and therefore admits the usual representation (expected utility, exponential discounting, etc.); and (iii) the sup‑norm distance between u and v is bounded by a function δ(ε) that goes to zero as ε → 0. In symbols, ‖u – v‖∞ ≤ δ(ε).

To obtain this result the authors draw on the theory of stability of functional equations (Hyers‑Ulam stability) and on Anderson’s 1986 “almost implies near” principle. The paper applies the general theorem to three canonical domains: (1) choices under risk, where the key axiom is the Independence axiom; (2) choices under uncertainty, using a streamlined Anscombe‑Aumann framework in which linearity (or state‑wise expected utility) is the pivotal property; and (3) intertemporal choice, where a stationary (time‑invariance) axiom is central. In each case they define an ε‑relaxed version of the axiom, prove the existence of a “nearby” standard utility, and give explicit bounds on δ as a function of ε.

In the risk domain, the authors show that an ε‑independence condition guarantees a utility u that rationalizes the observed lottery preferences and a standard expected‑utility function v that are δ‑close. The bound δ(ε) is linear in ε under the usual continuity, monotonicity, and extremality assumptions. In the uncertainty domain, they prove analogous results for linear (state‑wise) utilities and for homothetic preferences; when homotheticity holds the linear representation becomes exact (Theorem 3). They illustrate the theory with the max‑min expected‑utility model (which cannot be approximated by a subjective EU) and with the smooth ambiguity model, showing that when ambiguity neutrality fades for large stakes a subjective EU approximation exists.

For intertemporal choice, two treatments are presented—discrete‑time and continuous‑time. A relaxed stationary axiom yields a utility that is close to an exponential discounting function. Moreover, if the discount factors are strictly decreasing, the relaxed axiom actually forces an exact geometric discounting representation (Theorem 10).

A concrete empirical illustration is provided using the classic Allais paradox. The authors note that the observed pattern violates Independence by roughly ε ≈ 0.02. By plugging this ε into their bound they find that the expected‑utility representation and a cumulative‑prospect‑theory representation differ by at most 0.1 in normalized utility, confirming that the two utilities are indeed “near”. This demonstrates how a small measured deviation from an axiom can rationalize the use of a standard model as an approximation.

Beyond the three domains, the paper establishes a conceptual equivalence between two quantitative exercises that are common in behavioral economics: (a) measuring the size of an axiom violation (ε) as a test of rationality, and (b) assuming that observed choices are ε‑optimal solutions to the theoretical optimization problem. The authors show that the distance δ between the behavioral utility and the theoretical utility provides a bridge: a small ε implies a small δ, which in turn justifies treating the observed behavior as an ε‑optimal solution of the standard model. This connection gives a rigorous foundation to the “approximate optimization” approach popularized by Simon (1955) and widely used in computer science and game theory.

The literature review positions the work relative to recent papers that relax first‑order optimality conditions (de Clippe & Rozen 2023; Echenique et al. 2023) and to the classic Afriat‑Houtman‑Maks rationalizability literature. The novelty lies in starting from the axiom itself, rather than from the optimization problem, and in providing a general methodological framework that can be applied across decision‑theoretic domains.

In summary, the paper makes three substantive contributions: (1) it formalizes the “almost implies near” phenomenon for core decision‑theoretic axioms; (2) it derives explicit, testable bounds linking empirical axiom violations to the proximity of standard utility representations; and (3) it bridges the gap between empirical rationality tests and approximate optimization, thereby offering a unified justification for using standard models as approximations when observed behavior only approximately satisfies the underlying axioms.